Skip to content

EvoAgent: An Evolvable Agent Framework with Skill Learning and Multi-Agent Delegation

🕒 Published (v1): 2026-04-22 03:00 UTC · Source: Arxiv · link

Why this paper was selected

Structured skill learning with hierarchical sub-agent delegation in EvoAgent framework

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

EvoAgent is an evolvable LLM agent framework that couples structured multi-file skill packages with a hierarchical sub-agent delegation mechanism, evolving through a user-feedback-driven dual-loop (online execution + offline evolution). Applied to foreign trade scenarios, integrating EvoAgent with GPT-5.2 yields ~28% improvement in overall LLM-as-Judge scores across five evaluation dimensions.

Problem

Existing skill-based agents rely on manual skill authoring (high cost, poor scalability) and exhibit human–AI cognitive misalignment when human-intuitive workflows conflict with LLM reasoning patterns. Self-evolution methods target single tools or prompts rather than structured multi-file skill packages, and require labeled supervision or expert feedback unavailable in real-world deployments. Multi-agent frameworks use static role definitions with no adaptive delegation logic that evolves from historical experience.

Method

EvoAgent organizes skills as six-tuple structured packages \(\text{skill} = (\text{name}, \text{desc}, \text{triggers}, \text{instr}, \text{refs}, \text{meta})\), where \(\text{meta} = (\text{usage\_count}, \text{success\_rate}, \text{created\_at}, \text{updated\_at})\). The system models its task environment as an MDP \(\mathcal{M}_\text{EA} = (S, A, P, R)\) with state \(s = (h, u, S_\text{skills}, c)\) (dialogue history, user profile, skill repository, compression state) and reward: $\(R(s, a, s') = w_1 \cdot \text{Maturity}(\text{skill}) + w_2 \cdot \Delta u_\text{profile} + w_3 \cdot \Delta \text{memory}\)$ with optimization objective \(\max \mathbb{E}[\sum_{t=1}^T \gamma^t R(s^{(t)}, a^{(t)}, s^{(t+1)})]\).

Skill retrieval uses a three-stage cascade: (1) keyword trigger matching, (2) cosine-similarity embedding matching \(M_\text{embed} = \arg\max_{\text{skill}} \cos(E(u), E(\text{skill.desc}))\), (3) LLM-based intent classification as fallback. The online loop handles skill injection, context assembly, and dynamic history compression; the offline loop asynchronously evaluates sessions, updates a three-layer memory (SOUL.md/USER.md/MEMORY.md), and categorizes skills into four maturity tiers (Budding → Growing → Mature → Proficient) for pruning/promotion decisions. A hierarchical sub-agent delegation architecture lets the main agent spawn specialized sub-agents with independent context spaces on demand.

Key Contributions

  • Structured multi-file skill representation with lazy-loading references and evolutionary metadata tracking usage frequency and success rate.
  • User-in-the-loop closed-loop skill evolution that removes dependence on labeled data or expert annotation.
  • Three-stage cascaded skill matching (keyword → embedding → LLM) balancing retrieval cost and accuracy.
  • Hierarchical sub-agent delegation architecture enabling dynamic spawning of context-isolated specialist agents.
  • Three-layer persistent memory system (SOUL.md / USER.md / MEMORY.md) with dialogue compression preserving structured asset indices.
  • Online/offline dual-loop design explicitly separating real-time deterministic execution from asynchronous capability evolution, grounded in the Harness Engineering paradigm.

Results

  • GPT-5.2 with EvoAgent achieves ~28% improvement in overall average score under a five-dimensional LLM-as-Judge evaluation protocol on real-world foreign trade tasks (professionalism, accuracy, practical utility are cited dimensions; baseline is GPT-5.2 without EvoAgent).
  • Model transfer experiments show performance depends jointly on the base model's intrinsic capability and the degree of synergy between model and agent architecture (no additional numerical breakdowns provided in the excerpt).

Limitations

  • Evaluation confined to a single domain (foreign trade / B2B scenarios at Focus Technology); generalizability to other professional domains is undemonstrated.
  • The MDP reward function is theoretical; in practice, no explicit numerical reward is computed — the framework relies on implicit proxies (usage count, success rate), whose alignment with actual task quality is unverified.
  • LLM-as-Judge evaluation is subject to well-documented biases; no human evaluation or held-out test set results are reported in the provided text.
  • Offline evolution loop latency and its impact on session throughput are not quantified.
  • The three-stage skill matching cascade's sensitivity to embedding quality and threshold selection is not ablated.

Relevance to Harnesses / Meta-Harnesses

EvoAgent is a direct instantiation of Harness Engineering: it explicitly cites the Hashimoto/OpenAI/Böckeler–Fowler/Iusztin lineage and operationalizes the three pillars (Constraints, Observability, Feedback Loops) as concrete architectural components. The online/offline dual-loop separation mirrors the meta-harness pattern of separating a deterministic execution shell from an asynchronous evolution controller — the harness governs the model rather than being governed by it. The structured skill package (multi-file, versioned, metadata-tracked) and the three-layer memory system are engineering-grade harness components directly relevant to anyone designing harnesses that must accumulate and manage evolving capabilities. The MDP formalization of harness state and the maturity-tiered skill lifecycle provide a reusable theoretical vocabulary for the harness/meta-harness research thread.